Masked Discrete Diffusion Models (MDMs) enable parallel generation via iterative refinement. However, we identify a critical decisional mismatch. The MDM architecture is inherently dynamic and capable of sensing context shifts. In contrast, prevailing decoding paradigms remain static and myopic. They treat each denoising step as an isolated snapshot, effectively discarding valuable temporal feedback that signals logical conflicts. To bridge this gap, we propose Regret-Aware Confidence Calibration (RACC). This training-free framework aligns decoding decisions with the model’s latent self-correction capabilities. RACC introduces a momentum anchor to track confidence trajectories. When a token’s probability drops abruptly below its historical trend, the system triggers a “regret” signal. Unlike expensive re-masking or lookahead search, RACC utilizes this signal to proactively demote unstable candidates. Extensive experiments on reasoning benchmarks, such as HumanEval and GSM8K, demonstrate that RACC significantly improves generation consistency. Crucially, RACC achieves these gains with zero additional inference overhead, effectively balancing decoding quality and efficiency.

Narrow-domain finetuning makes vision-language models (VLMs) sound like domain experts while quietly making them less attentive to the image. We attribute this gap to a persistent domain prior: an input-agnostic shift in hidden state, installed by narrow finetuning, that biases the model toward domain-style outputs regardless of visual evidence. To test the claim we introduce ActDiff, a white-box diagnostic that isolates the finetuning trace by differencing base and finetuned hidden states under progressively degraded images. The trace is compact and persists under blank and Gaussian-noise inputs; the finetuned model is correspondingly less, not more, disrupted when the image is destroyed (gap 0.323 ± 0.036 nats on BIOMED, generic eval). Subtracting the trace at inference recovers 54% of the MMStar accuracy collapse on BIOMED, with a peak-and-collapse dose response that random and wrong-domain controls do not produce. A single first-moment regulariser added at training time shrinks the trace 34x and improves generic NLL. The prior thus has a clean diagnostic, a causal handle at inference, and a one-line mitigation at training.